[elbruno]

Summary

Closes #58.

• Validate actual Laya Choice, Score and Assessment through the shared ElBruno.AI.Decisions.IDecisionClient, including contrastive cases, full probability invariants, local HTTP isolation, calibration diagnostics and a documented model-quality counterexample.
• Add a separate KevOnnxDecisionClient and AddLocalKevDecisions() backend over explicitly acquired CPU FP32 ONNX assets. Laya stays the default; Ollama is only the model-discovery catalog, not a backend or conversion input.
• Convert the pinned original Apache-2.0 Kev 0.8B adapter/base and trained pointer head with exact original temperature, strict context limits, manifest/provenance and mandatory PyTorch reference parity. Preserve exact free-text tokenization, semantic labels and all option probabilities.
• Record Tev1's unresolved weight license and the remaining model families' distinct head/export contracts. No unexecuted model is advertised as supported.
• Prepare coordinated NuGet release 0.24.0 and require ONNX Runtime 1.31.0 for Kev's linear-attention operator.

Verification

• Re-ran the repository Kev converter end-to-end, then all three shared operations from free text in .NET. Maximum probability difference against original FP32 PyTorch: approximately 5.614e-6, below 1e-4.
• Exact token/marker parity, Unicode/special-token cases, cancellation recovery, concurrent callers and no additional HTTP attempts during local loading/inference.
• 99 decision tests passed with both live models enabled; 97 pass and two explicitly skip without opt-in.
• All five publishing test suites passed locally: 1,947 tests total.
• Solution Release build passed with the repository CI's existing -p:NoBuild=false override.
• All five NuGet packages packed and passed package/assembly version validation for 0.24.0.
• SharedDecisions ran on the converted Kev export, including decision routing.

Scope and limitations

Only Kev 0.8B text/CPU/FP32 is validated. Conversion uses Python once; application runtime needs neither Python nor a server. Weights are not included in Git or NuGet. See docs/kev-guide.md and docs/decision-model-validation.md. Tev1 remains unimplemented and blocked on its weight license; six synthetic cases are not an accuracy or calibration benchmark.